M1: Config and OpenAI-compatible provider client

A pydantic config object and an httpx client that speaks /v1/chat/completions against a configurable base URL, with named presets for LM Studio (127.0.0.1:1234), Ollama (127.0.0.1:11434) and llama.cpp (127.0.0.1:8888), each overridable by environment variable. Local servers ignore the API key but usually require one to be present, so send a placeholder.

Timeouts matter more here than for cloud providers: a 4B model on a loaded laptop can take a long time to first token, and a default 30s timeout will produce failures that look like harness bugs.

Disable proxy environment discovery on the client (trust_env=False in httpx, or an explicit no-proxy for loopback). With ALL_PROXY or HTTPS_PROXY set, httpx routes even 127.0.0.1 through the proxy and the failure surfaces as "LM Studio isn't responding" rather than as a proxy error. Observed for real while probing the sandbox; see ADR-0005.

Done when a smoke script gets a completion back from a running LM Studio server, and the client's request construction is unit-tested against a recorded response with no server running.

Start with ADR-0005 and the /v1/chat/completions contract described in the issue. Trace the project's existing configuration and client entry points before defining the presets, environment overrides, timeout, and proxy behavior. Done means a smoke script receives a completion from LM Studio and unit tests verify request construction against a recorded response without a running server.